1 / 33

Speech Recognition

Speech Recognition. Training Continuous Density HMMs Lecture Based on: Dr. Rita Singh’s Notes School of Computer Science Carnegie Mellon University. Table of Content. Review of continuous density HMMs Training context independent sub-word units Outline Viterbi training

emilia
Download Presentation

Speech Recognition

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Speech Recognition Training Continuous Density HMMs Lecture Based on: Dr. Rita Singh’s Notes School of Computer Science Carnegie Mellon University

  2. Table of Content • Review of continuous density HMMs • Training context independent sub-word units • Outline • Viterbi training • Baum-Welch training • Training context dependent sub-word units • State tying • Baum-Welch for shared parameters Veton Këpuska

  3. Discrete HMM • Data can take only a finite set of values • Balls from an urn • The faces of a dice • Values from a codebook • The state output distribution of any state is a normalized histogram • Every state has its own distribution HMM outputs one of four colors at each instant. Each of the three states has a different probability distribution for the colors. Since the number of colors is discrete, the state output distributions are multinomial. Veton Këpuska

  4. Continuous Density HMM • There data can take a continuum of values: • e.g., cepstral vectors • Each state has a state output density. • When the process visits a state, it draws from the state output density for that state. HMM outputs a continuous valued random variable at each state. State output densities are mixtures of Gaussians. The output at each state is drawn from this mixture. Veton Këpuska

  5. Modeling Output State Densities • The state output distributions might be anything in reality • We model these state output distributions using various simple densities • The models are chosen such that their parameters can be easily estimated • Gaussian • Mixture Gaussian • Other exponential densities • If the density model is inappropriate for the data, the HMM will be a poor statistical model • Gaussians are poor models for the distribution of power spectra Veton Këpuska

  6. Insufficient data to estimate all parameters of all Gaussians Assume states from different HMMs have the same state output distribution Tied-state HMMs Assume all states have different mixtures of the same Gaussians Semi-continuous HMMs Assume all states have different mixtures of the same Gaussians and some states have the same mixtures Semi-continuous HMMs with tied states Other combinations are possible Parameter Sharing Veton Këpuska

  7. Training Models for a Sound Unit • Training involves grouping data from sub-word units followed by parameter estimation. AC AX F AC K S IH N S AC K S AC N B AC K S AC N N AC K S FOX IN SOCKS ON BOX ON KNOX Veton Këpuska

  8. Training involves grouping data from sub-word units followed by parameter estimation Indiscriminate grouping of vectors of a unit from different locations in the corpus results in Context-Independent (CI) models Explicit boundaries (segmentation) of sub-word units not available We do not know where each sub-word unit begins or ends Boundaries must be estimated Training models for a sound unit For a 5-state HMM, segment data from each instance of sub-word unit to 5 parts, aggregate all data from corresponding parts, and find the statistical parameters of each of the aggregates Veton Këpuska

  9. Learning HMM Parameters • Viterbi training • Segmental K-Means algorithm • Every data point associated with only one state • Baum-Welch • Expectation Maximization algorithm • Every data point associated with every state, with a probability • A (data point, probability) pair is associated with each state Veton Këpuska

  10. Viterbi Training • Initialize all HMM parameters • For each training utterance, find best state sequence using Viterbi algorithm • Bin each data vector of utterance into the bin corresponding to its state according to the best state sequence • Update counts of data vectors in each state and number of transitions out of each state • Re-estimate HMM parameters • State output density parameters • Transition matrices • Initial state probabilities • If the likelihoods have not converged, return to step 2. Veton Këpuska

  11. Viterbi Training: Estimating Model Parameters • Initial State Probability • Initial state probability (s) for any state s is the ratio of the number of utterances for which the state sequence began with s to the total number of utterances. Veton Këpuska

  12. Viterbi Training: Estimating Model Parameters • Transition probabilities • The transition probability a(s,s’) of transiting from state s to s’ is the ratio of the number of observations from state s, for which the subsequent observation was from state s’, to the number of observations that were in s Veton Këpuska

  13. Viterbi Training: Estimating Model Parameters • State output density parameters • Use all the vectors in the bin for a state to compute its state output density • For Gaussian state output densities only the means and variances of the bins need be computed • For Gaussian mixtures, iterative EM estimation of parameters is required within each Viterbi iteration Posterior probability of kth Gaussian given observation x. P(x|k) is a Gaussian. Veton Këpuska

  14. Viterbi Training: Estimating Model Parameters • Mean of kth Gaussian • Covariance of kth Gaussian • Mixture weight of kth Gaussian Veton Këpuska

  15. Baum-Welch Training • Initialize HMM parameters • On each utterance run forward backward to compute following terms: • utt(s,t) = a posterioriprobability given the utterance, that the process was in state s at time t • utt(s,t,s’,t+1) = a posterioriprobability given the utterance, that the process was in state s at time t, and subsequently in state s’ at time t+1 • Re-estimate HMM parameters using gamma terms • If the likelihood of the training set has not converged, return to step 2. Veton Këpuska

  16. Baum-Welch: Computing A Posteriori State Probabilities and Other Counts • Compute  and  terms using the forward backward algorithm Veton Këpuska

  17. Baum-Welch: Computing A Posteriori State Probabilities and Other Counts • Compute a posteriori probabilities of states and state transitions using  and  values Veton Këpuska

  18. Baum-Welch: Estimating Model Parameters • Initial State Probability • Initial state probability (s) for any state s is the ratio of the expectednumber of utterances for which the state sequence began with s to the total number of utterances Veton Këpuska

  19. Baum-Welch: Estimating Model Parameters • Transition probabilities • The transition probability a(s,s’) of transiting from state s to s’ is the ratio of the expectednumber of observations from state s for which the subsequent observation was from state s’, to the expectednumber of observations that were in s Veton Këpuska

  20. Baum-Welch: Estimating Model Parameters • State output density parameters • The a posteriori state probabilities are used along with a posteriori probabilities of the Gaussians as weights for the vectors • Means, covariances and mixture weights are computed from the weighted vectors Posterior probability of kth Gaussian given observation x. P(x|k) is a Gaussian. Veton Këpuska

  21. Baum-Welch: Estimating Model Parameters • Mean of kth Gaussianof state s • Covariance of kth Gaussianof state s • Mixture weight of kth Gaussianof state s Veton Këpuska

  22. Context based grouping of observations results in finer, Context-Dependent (CD) models CD models can be trained just like CI models, if no parameter sharing is performed Usually insufficient training data to learn all triphone models properly Parameter estimation problems Parameter estimation problems for CD models can be reduced by parameter sharing. For HMMs this is done by cross-triphone, within-state grouping Training Context Dependent (Triphone) Models Veton Këpuska

  23. The assignment of vectors to states can be done using previously trained CI models or with CD models that have been trained without parameter sharing Partitioning any set of observation vectors into two groups increases the average (expected) likelihood of the vectors The expected log-likelihood of a vector drawn from a Gaussian distribution with mean  and variance C is Grouping of context-dependent units for parameter estimation Veton Këpuska

  24. Expected log-likelihood of a vector drawn from a Gaussian distribution • This is a function only of the variance of the Gaussian • The expected log-likelihood of a set of N vectors is Veton Këpuska

  25. If we partition a set of N vectors with mean  and variance C into two sets of vectors of size N1and N2, with means 1and 2and variances C1and C2respectively, the total expected log-likelihood of the vectors after splitting becomes Grouping of context-dependent units for parameter estimation • The total log-likelihood has increased by Veton Këpuska

  26. Observation vectors partitioned into groups to maximize within class likelihoods Recursively partition vectors into a complete tree Prune out leaves until desired number of leaves obtained The leaves represent tied states (sometimes called senones) All the states within a leaf share the same state distribution 2n-1 possible partitions for n vector groups. Exhaustive evaluation too expensive Linguistic questions used to reduce search space Grouping of context-dependent units for parameter estimation Veton Këpuska

  27. Linguistic Questions • Linguistic questions are pre-defined phone classes. Candidate partitions are based on whether a context belongs to the phone class or not • Linguistic question based clustering also permits us to compose HMMs for triphones that were never seen during training (unseen triphones) Veton Këpuska

  28. Composing HMMs for unseen triphones • For every state of the N-state HMM for the unseen triphone, locate appropriate leaf of the tree for that state • Locate leaf by answering the partitioning questions at every branching of the tree Vowel? S or Z? Veton Këpuska

  29. Linguistic Questions • Linguistic questions are pre-defined phone classes. Candidate partitions are based on whether a context belongs to the phone class or not • Linguistic question based clustering also permits us to compose HMMs for triphones that were never seen during training (unseen triphones) • Linguistic questions must be meaningful in order to deal effectively with unseen triphones • Meaningful Linguistic Questions? • Left Context: (A,E,I,Z,SH) • ML Partition: (A,E,I) (Z,SH) • (A,E,I) vs. Not (A,E,I) • (A,E,I,O,U) vs Not (A,E,I,O,U) I A Z E SH • Linguistic questions can be automatically designed by clustering of context-independent models. Veton Këpuska

  30. Other forms of parameter sharing • Ad-hoc sharing: sharing based on human decision • Semi-continuous HMMs – all state densities share the same Gaussians • This sort of parameter sharing can coexist with the more refined sharing described earlier. Veton Këpuska

  31. Baum-Welch: Sharing Model Parameters • Model parameters are shared between sets of states • Update formulae are the same as before, except that the numerator and denominator for any parameter are also aggregated over all the states that share the parameter Veton Këpuska

  32. Mean of kth Gaussian of any state in the set of states  that share the kth Gaussian Covariance of kth Gaussian of any state in the set of states that share the kth Gaussian Mixture weights of kth Gaussian of any state in the set of states that share the kth Gaussian mixture Baum-Welch: Sharing Model Parameters Veton Këpuska

  33. Conclusions • Continuous density HMMs can be trained with data that have a continuum of values • To reduce parameter estimation problems, state distributions or densities are shared • Parameter sharing has to be done in such a way that discrimination between sounds is not lost, and new sounds are accounted for • Done through regression trees • HMMs parameters can be estimated using either Viterbi or Baum-Welch training Veton Këpuska

More Related